The O'Reilly 2nd Annual AI Adoption in the Enterprise Report reveals all the key trends shaping the rapid growth of AI, including:
- Among mature adopters, supervised learning was reported to be the most popular machine learning technique (73%), while deep learning (55%) is the most popular among organizations still in the evaluation stage of artificial intelligence.
- Difficulties in hiring and retaining people with AI skills was once again noted as a top barrier to AI adoption in the enterprise.
- More than 26% of respondents say their organizations plan to institute formal data governance processes and/or tools by 2021 and nearly 35% expect this to happen in the next three years
The Current State of AI in the Enterprise
Roumeliotis: "I think that there's still long ways to go for companies. I mean, every company is a software company is what they say, right? And so I think that's why you see a lot in research and development, I think that people are trying a lot of different things. I think IT is something where you're gonna hear, you're gonna start to hear the term AIOps a lot.
“So specifically, AIOps has come to mean automation of observability of deployment, metrics, stuff like that that can scale up, can scale down, resiliency, security. The insights that AI can provide beyond what we could easily see. So that's I think the next thing that you're gonna see really explode with AI.
“So there's that, and then customer service – I think I'm not surprised that comes up as well. We've talked about how you can get a tech spot at a hotel that says, "If you need anything, let me know," and then you can say, "Hey, I need towels. Bring them to my room." Or Alexa, for instance. Everybody knows Alexa.
Barriers To AI Adoption
Roumeliotis: “I think there's a lot of people that mis-characterize AI, like I said. AIOps is commonplace in five years, I think at the outlier, right?
"And people aren't gonna necessarily think about that as AI. I think chatbots and language processing and computer vision, these are the easier types of things to think of as AI. Those aren't necessarily the use cases that make sense for everybody, but enhancing business insights through analytics that use AI, that's something that I think you're gonna see a lot about too. That idea of the business analyst moving from going through Excel sheets to really being able to do some basic scripting or having the software do scripting for him or her.
“But yeah, it's definitely completely new. So it's something that you have to think about, but I also think you have to open your mind to what AI is, and what we've seen is that for most people who are at the beginning of that journey, they're interested in finding out what other companies are doing. And so we have a lot of case studies.
Maguire: The thing that also caught my eye there was: lack of skilled people, difficulty hiring the required roles. I definitely hear that skilled AI people are really commanding pretty lofty salaries, to the idea of hiring an in-house AI expert is a seriously expensive proposition.
Roumeliotis: “Yeah, it's specialty upon specialty, right? So a data scientist, already a specialist, a data scientist is now also becoming more of a hybrid between a data scientist and a data engineer, figuring out how to do all of that work and figuring that out. But also figuring out how it works and how you move that data around into the cloud, as you mentioned.
“And so the idea that you then need another person, an AI engineer to build out algorithms and figure out how that's gonna work for you is another expense. I do think in the future that we will see data markets and that will include the ability to buy algorithms as well, because not everything is gonna need something super custom, so I think that's coming.
Maguire: So this is interesting: What kinds of risks do you check for during machine learning, model building and deployment. I thought it was interesting that people said fairness, bias and ethics, I'm heartened to see that score is high, I'm almost surprised that it is as high as this survey suggests it is.
Roumeliotis: “So I think that they're the reason. Well, one, like you said, I think it's good that people care, so that's good, right?
“But I think, if I was gonna put on my pessimist's hat, I think that the reason people care is that when things go wrong, they go really wrong and we hear about them. There was some AI chat bot on Twitter that was just terrible and racist and awful and so that garnered a lot of terrible news coverage, as it should have.
“So I think that hopefully you wanna do it because you care, and even if you don't care, then you wanna do it because you care about yourself and your company. Yeah, I would think that you want to have that clear insight, right? There's so many different types of bias and that's not gonna help your company ultimately either, right?
“So AI regulatory insights is gonna be something [big]. Whether we should have them, whether we shouldn't. But that is the perfect example, right? We've decided yes or no for you to get this loan. You have to be able to tell the person why. You could think about this in healthcare, you could think about this in the supply chain, there has to be that transparency, absolutely. They go side by side, really, the fairness, and the transparency.
“One thing, if you can talk about AI being everyday, is the unexpected outcomes predictions. I think we're just starting to see that, like a lot of people, like I said, the Alexas are out there, and there's the story about how people had a conversation about a certain person, and then they sent that 60-minute conversation to the person, so that's an unexpected outcome.
“But this idea that that first wave of sort of not Google are starting to put AI projects into production, and this idea of model drift, or algorithmic accountability, I think dealing with that practically when you don't know exactly what that's gonna look like, and there's not a lot of sort of insight into what that's gonna look like. I'm not surprised that that is something that people are most worried about. In fact, we are doing a lot of content on our O'Reilly learning platform, about that.
“That is sort of like, so many people have come to us about privacy issues, security governance. I know we're probably gonna talk about, and figuring out not only how do you keep it on track, but what do you do when it goes off track.
“So you do the best that you can, you train it, there's supervised learning, but then ultimately, you're right, some of the more advanced stuff is unsupervised, and will continue to learn and adapt and you almost need another AI, piece of AI to then go in and observe it and follow it. Right, but, yeah, there's a lot, like we were saying, the dumber AI is sort of human in the loop or supervised learning, but that's just the beginning.
Viewer Question: How will the the COVID-19 crisis affect AI? Plus: Deep Learning and Data Governance
Roumeliotis: “I think that we're seeing for the first time in the mainstream news, modeling, the data modeling and a lot of that is supported by, being enhanced by AI. And I think we're seeing the good and the bad with that, we're seeing, honestly, how important the data is that people have to those models.
“So AI really isn't anything without all of that data and we're at such the beginning of, unfortunately, this crisis that... The data is just piling up every day. So the models that we had four to six weeks ago said one thing. Literally today, they say something else.
“So I think it's showing people…a true picture of what AI can do as far as predictive modeling and I think it is giving people probably ideas about how they could use that in their own business. As far as healthcare specifically, I would love to see more AI predictive modeling in general, and being able to somehow use anonymized data to figure things out. And I know it's a business like anything else and it costs money, so I can understand why it hasn't happened, but the thought of what we could do if we synthesized all that data is stunning.
Maguire: Deep learning continues to rise. I found this really again quite surprising – deep learning is a very sophisticated. must say, I'm pretty surprised.
Roumeliotis: “So we talked a little bit about this previously too, like this idea that this is probably a little bit skewed towards our audience and the software industry. Again, I think it's sort of this conversational commerce, it's the idea. It's like neural nets, which is pretty high tech, and neural nets have been around for a long time, but this idea of the natural language processing, interacting with voice, the chat bots are used in there.
“Computer vision, I feel like we see on the periphery of things, but we could see more of that. So I think that's where you're seeing that come in, and obviously that can be either supervised or unsupervised or transfer and stuff like that, but, yeah. And we're seeing that reflected in our learning platform as well. That is definitely what people are looking at right now. The one that's coming up sort of next that we're seeing is reinforcement learning.
“So that I can see where human in a loop industrial, where there needs to be another set of eyes, mission critical, that sort of thing. And I know there are surgeries where doctors are doing the surgery through a machine, and maybe we'll get to a point where that's what happens, that machines do surgery, but I for one would feel good to have the steadiness of the machine possibly doing the cutting, but the brains of the doctor still in the loop.
“So TensorFlow from Google really is sort of the gateway that basically open source opened the AI world I think much faster than would have happened had there been no giant open source framework. So that's super important. And you'll notice a lot of these are Python-based and so Python is approachable and a lot of people have used that in data, and it's actually a programming language that a lot of people use, even when they're first joining, going into computer science in general.
“So PyTorch is one that's been coming up. It's been most used in academia and has had issues with production, but they're working on that, and that is the tool that I would watch other than TensorFlow. Yeah, and is in some ways, better than TensorFlow. But that's the one to watch, Keras is still really popular. And then we sort of get down into specific cloud. I think the cloud stuff you'll see come up more and more, like we talked about sort of like this idea of self-service AI.
Data governance and AI
Roumeliotis: “I think that data governance is a touchy area in that I think is it something that we all govern ourselves, we all look at everything or should there be some sort of ultimate legislation. But this idea of, on many levels it's important, I think. And the idea of data lineage and security and privacy and we talked about ethics and bias. So it's something that is thankfully coming up. It's sort of like the vegetables of AI - so I think it's not the most fun part of putting together a project, but you need to think about it and you need to track it."
The Future of AI
Roumeliotis: “I think that we are gonna see much more simple at first, self-service data solutions. They exist now, but I think they're gonna get more and more sophisticated and I think you're gonna see them either, I mean, we see them now, but see them from the cloud providers, the cloud providers, or we're gonna see companies where they play very well with the cloud providers.
“And that is because it's answering that need of expertise that is hard to come by. I think that people are understanding that data quality is probably, again, a vegetable-type thing. Where not super fun, but de-duping and making sure that you're pulling in clean data. That becomes more and more important. And then I think that people are gonna get more savvy in their understanding of AI and we're gonna see it pop up more and more. And sometimes you're not even gonna realize it.
“Like most technologies that start out specialized, I think [AI] is gonna become more and more democratized. And I think that need for specialists will never go away, but I think that more people will be able to do it. The other thing is that I think with machine learning, not every project is gonna need terabytes of data in all this stuff. A lot of base level stuff is just gonna need sort of like your regular CPU and nothing too crazy. And the answers are gonna be just good enough, that they're fine. So I think we're gonna see more AI: like we said, the simple pieces of AI, I think we'll see all over the place."